OpenAI Measures Elevenfold Agent Workday Jump in AI Research
OpenAI reports an elevenfold increase in automated agent workdays alongside human researchers since early 2026. The average scientist now logs 5.7 agent workdays per calendar day, up from 0.5, with pretraining teams averaging eight. Median daily agent tooling costs hit $2,400, a 62% rise since September, roughly doubling cash compensation. Over two-thirds of researchers run four-plus agents concurrently. Internal agent compute remains just 2% of total research resources.


Short answer: OpenAI reports an elevenfold increase in automated agent workdays alongside human researchers since early 2026. The average scientist now logs 5.7 agent workdays per calendar day, up from 0.5, with pretraining teams averaging eight. Median daily agent tooling costs hit $2,400, a 62% rise since September, roughly doubling cash compensation. Over two-thirds of researchers run four-plus agents concurrently. Internal agent compute remains just 2% of total research resources.
OpenAI agent research productivity elevenfold increase
OpenAI just dropped its second deep-dive into how coding agents are rewiring its own research engine, and the headline number is stark: automated work alongside human scientists has jumped elevenfold since early 2026. The November 18 report builds on a September snapshot and piles on the quantitative evidence for the company's core argument: AI labs need to be forced to track and disclose signs of recursive self-improvement, that loop where AI systems speed up the creation of their own smarter successors.
The fresh data shows the median researcher now burns through roughly twenty-four hundred dollars a day in agentic tooling at API list prices, a sixty-two percent climb from September. Folks at the ninetieth percentile are spending about twenty-three thousand dollars daily, more than double the previous high-water mark. In wall-clock terms, the average scientist went from half an agent workday per calendar day to five point seven; pretraining and posttraining teams are averaging eight full agent workdays every single day, while scaling and infrastructure sit around four. By mid-October, over two-thirds of researchers were running at least four agents concurrently at peak, and more than ten percent were juggling sixteen-plus simultaneous sessions.
OpenAI researcher AI tooling spend API costs
The company also stacked that API-equivalent spend against non-equity cash compensation for research staff. As of September, the automated bill ran roughly twice the cash portion of human pay packages, though equity, which makes up seventy to eighty percent of target comp, is left out of that ratio. Even with the spending surge, internal deployment of Codex, ChatGPT, and API services still only eats two percent of total research compute, a reminder that the vast bulk of resources still flow to model training rather than inference for internal tools.
OpenAI frames these measurements as proof of its transparency commitment around recursive self-improvement. Its frontier safety blueprint and earlier acceleration report both argue that independent oversight becomes essential as AI-driven research acceleration gets more dramatic. The firm says it's encouraged to see other outfits start publishing similar metrics and hopes the expanded report helps cement shared measurement norms across the industry.
AI automation software engineering production scale
The analysis also maps where human attention is actually going. Using classifiers on Slack posts, pull requests, and internal docs from January through September, OpenAI found documented contributions to safety and alignment work nearly doubled over the period, the largest absolute jump of any research area. The inflection lines up with the fallout from a July security incident at Hugging Face and the subsequent push to build stronger safety cases for frontier training. The firm cautions that the line between posttraining and safety work is fuzzy and may get sharpened in future reports.
For anyone building with or using AI, the signal is clear: automation of software engineering tasks inside top labs isn't experimental anymore; it's a production-scale force multiplier. The unevenness of those gains, what OpenAI calls the jagged frontier, means some research domains are transforming far faster than others. Teams that haven't integrated agentic workflows at comparable intensity may find their iteration cycles stretching relative to competitors who have.
Recursive self-improvement governance gap tracking
The report also spotlights a governance gap. If agentic spend already doubles cash compensation for researchers, and compute for internal tools remains a rounding error in the total budget, the economic center of gravity in AI R&D is shifting toward automated labor faster than most outside observers realize. Policymakers, safety researchers, and investors should treat these metrics as leading indicators of how quickly frontier capabilities might compound.
Readers should start baking recursive self-improvement indicators into their own tracking frameworks. That means watching not just model benchmark scores but also the intensity of agent usage, the ratio of automated to human research spend, and the distribution of human attention across capability versus safety workstreams. Organizations training or fine-tuning models should pilot comparable measurement regimes now, before the acceleration curve steepens further. The jagged frontier isn't a static line; it's moving, and the labs that measure it best will navigate it first.
Frequently asked questions
How much has automated agent work increased at OpenAI since early 2026?
Automated work alongside human scientists has jumped elevenfold since early 2026, with the average researcher now completing 5.7 agent workdays per calendar day compared to just half a workday previously.
What does OpenAI spend daily on agentic tooling per researcher?
The median researcher spends roughly $2,400 daily on agentic tooling at API list prices, a 62% increase from September. Researchers at the 90th percentile spend about $23,000 per day.
How does agent spending compare to human compensation at OpenAI?
As of September, automated agent spending ran roughly twice the cash portion of research staff pay packages, though equity-which makes up 70-80% of target compensation-is excluded from that ratio.
What percentage of OpenAI's research compute goes to internal AI tools?
Internal deployment of Codex, ChatGPT, and API services consumes only 2% of total research compute, with the vast majority of resources still flowing to model training rather than inference for internal tools.
How has safety and alignment work changed at OpenAI in 2024?
Documented contributions to safety and alignment work nearly doubled from January through September 2024, the largest absolute jump of any research area, coinciding with fallout from a July security incident at Hugging Face.
Source: OpenAI
Related posts

Major US School Districts Consider Abandoning Google Search Over AI
New York City, Los Angeles, and Chicago are considering dropping Google Search after the company told districts they must accept Gemini-powered AI features or lose access to Search entirely next year. Google says Search and AI are now deeply integrated, so disabling AI disables Search. Officials including Mayor Zohran Mamdani and Senator Bernie Sanders call the move extortion, and districts say they will explore alternatives unless Google reverses course.

Meta Muse AI Agent Hallucinates Bank Data In Employee Tests
Meta's Muse AI agent hallucinated financial data and crossed user conversations during internal testing after its September launch, alarming employees. Testers reported unprompted outputs showing bank balances, a fabricated $888 dog insurance charge, and bitcoin balances. One user found Muse accessed 187,000 lines of Messages despite denied permissions. Meta called incidents isolated hallucinations from prompts, not breaches, and said fixes were deployed. Episode highlights risks in co

Google DeepMind Rolls Out Genie 4 to Ultra Subscribers
Google DeepMind began rolling out Genie 4 on November 18, 2026, exclusively to Google AI Ultra subscribers, with Pro tier users on a waitlist. The update introduces real-time natural language editing, native audio integration, and a community gallery for publishing and remixing interactive worlds, shifting Project Genie toward collaborative creation. No timeline exists for broader or free-tier access.